US Lawmaker Claims Nvidia Assisted Chinese AI Firm DeepSeek With Military-Linked Models

A letter from Representative John Moolenaar (R-MI), chair of the House Select Committee on the Chinese Communist Party, alleges that Nvidia provided technical assistance to Chinese AI company DeepSeek to optimize its large language models, which were later adapted for military use by China's People's Liberation Army (PLA).

Key Claims:

1. Technical Collaboration: Nvidia engineers reportedly worked directly with DeepSeek to optimize the company's AI models for Nvidia's H100 and A100 GPUs
2. Military Adaptation: The letter claims these optimized models were subsequently modified for military applications including battlefield simulations and autonomous systems
3. Timeline: The alleged collaboration occurred between Q3 2023 and Q2 2024, prior to stricter US export controls

What's New:

• DeepSeek's Technical Stack: The Chinese firm has developed several open-source models including DeepSeek-R1 (a 7B parameter LLM) and DeepSeek-V2 (a vision-language model), which the letter claims were optimized through Nvidia's assistance
• Military Applications: Specific PLA use cases mentioned include automated threat assessment systems and drone swarm coordination algorithms
• Legal Gray Area: The collaboration reportedly occurred during a period when Nvidia was still legally exporting data center GPUs to China

Technical Context:

• Model Optimization: Nvidia's CUDA toolkit and TensorRT-LLM software are standard tools for optimizing AI models on Nvidia hardware, used by developers worldwide
• Military Adaptation Process: PLA engineers allegedly modified the base models through techniques like reinforcement learning from human feedback (RLHF) tuned to military objectives
• Hardware Constraints: Despite US export restrictions, Chinese entities continue accessing restricted chips through third-party distributors and cloud providers

Limitations and Unanswered Questions:

1. Evidence Gap: The letter provides no technical documentation or specific model versions demonstrating military adaptation
2. Commercial vs Military Use: DeepSeek's models have legitimate commercial applications in healthcare and finance, complicating attribution
3. Nvidia's Compliance: Nvidia maintains it has "fully complied" with export control regulations, with CEO Jensen Huang recently stating the company follows "all government directives"

Industry Implications:

• Export Control Enforcement: Highlights challenges in preventing dual-use technology transfer even after hardware restrictions
• Open Source Concerns: Raises questions about military adaptation of openly published model weights (DeepSeek released several models under Apache 2.0 license)
• Supply Chain Monitoring: Underscores need for better tracking of end-use applications for exported AI technologies

Relevant Context:

• DeepSeek recently announced plans to develop a multi-language search engine and agentic AI systems
• Nvidia's China-focused A800 and H800 chips were specifically designed to comply with 2022 export controls before being banned in 2023
• The US has blacklisted several Chinese AI firms including SenseTime for military ties

Expert Perspectives:

• CSET Analyst: "This highlights the porous nature of AI export controls - model optimization knowledge transfers through standard developer support channels"
• Stanford HAI Researcher: "We need forensic techniques to detect military fine-tuning of foundation models, similar to nuclear proliferation detection methods"
• Nvidia Spokesperson: "We provide equal technical support to all developers within compliance boundaries. End-use monitoring is the responsibility of regulators"

The allegations come as the Commerce Department prepares new rules governing AI model exports, potentially restricting open-source releases of certain model weights. DeepSeek has not responded to requests for comment.